X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fcompiler%2Fseqtran.lisp;h=2699964568f975ba43d318d8111878074fa44fdc;hb=1a60ff79067ec697c476185e0c79565dacf8c8c0;hp=5e7a97bb4061baff9c11fba5bd4077857bba6e7d;hpb=c713eb2b521b048ff2c927ec52b861787d289f85;p=sbcl.git diff --git a/src/compiler/seqtran.lisp b/src/compiler/seqtran.lisp index 5e7a97b..2699964 100644 --- a/src/compiler/seqtran.lisp +++ b/src/compiler/seqtran.lisp @@ -15,38 +15,43 @@ (defun mapfoo-transform (fn arglists accumulate take-car) (collect ((do-clauses) - (args-to-fn) - (tests)) + (args-to-fn) + (tests)) (let ((n-first (gensym))) (dolist (a (if accumulate - arglists - `(,n-first ,@(rest arglists)))) - (let ((v (gensym))) - (do-clauses `(,v ,a (cdr ,v))) - (tests `(endp ,v)) - (args-to-fn (if take-car `(car ,v) v)))) - - (let ((call `(funcall ,fn . ,(args-to-fn))) - (endtest `(or ,@(tests)))) - (ecase accumulate - (:nconc - (let ((temp (gensym)) - (map-result (gensym))) - `(let ((,map-result (list nil))) - (do-anonymous ((,temp ,map-result) . ,(do-clauses)) - (,endtest (cdr ,map-result)) - (setq ,temp (last (nconc ,temp ,call))))))) - (:list - (let ((temp (gensym)) - (map-result (gensym))) - `(let ((,map-result (list nil))) - (do-anonymous ((,temp ,map-result) . ,(do-clauses)) - (,endtest (cdr ,map-result)) - (rplacd ,temp (setq ,temp (list ,call))))))) - ((nil) - `(let ((,n-first ,(first arglists))) - (do-anonymous ,(do-clauses) - (,endtest ,n-first) ,call)))))))) + arglists + `(,n-first ,@(rest arglists)))) + (let ((v (gensym))) + (do-clauses `(,v ,a (cdr ,v))) + (tests `(endp ,v)) + (args-to-fn (if take-car `(car ,v) v)))) + + (let* ((fn-sym (gensym)) ; for ONCE-ONLY-ish purposes + (call `(funcall ,fn-sym . ,(args-to-fn))) + (endtest `(or ,@(tests)))) + (ecase accumulate + (:nconc + (let ((temp (gensym)) + (map-result (gensym))) + `(let ((,fn-sym ,fn) + (,map-result (list nil))) + (do-anonymous ((,temp ,map-result) . ,(do-clauses)) + (,endtest (cdr ,map-result)) + (setq ,temp (last (nconc ,temp ,call))))))) + (:list + (let ((temp (gensym)) + (map-result (gensym))) + `(let ((,fn-sym ,fn) + (,map-result (list nil))) + (do-anonymous ((,temp ,map-result) . ,(do-clauses)) + (,endtest (truly-the list (cdr ,map-result))) + (rplacd ,temp (setq ,temp (list ,call))))))) + ((nil) + `(let ((,fn-sym ,fn) + (,n-first ,(first arglists))) + (do-anonymous ,(do-clauses) + (,endtest (truly-the list ,n-first)) + ,call)))))))) (define-source-transform mapc (function list &rest more-lists) (mapfoo-transform function (cons list more-lists) nil t)) @@ -71,37 +76,99 @@ ;;; MAP is %MAP plus a check to make sure that any length specified in ;;; the result type matches the actual result. We also wrap it in a ;;; TRULY-THE for the most specific type we can determine. -(deftransform map ((result-type-arg fun &rest seqs) * * :node node) - (let* ((seq-names (make-gensym-list (length seqs))) - (bare `(%map result-type-arg fun ,@seq-names)) - (constant-result-type-arg-p (constant-continuation-p result-type-arg)) - ;; what we know about the type of the result. (Note that the - ;; "result type" argument is not necessarily the type of the - ;; result, since NIL means the result has NULL type.) - (result-type (if (not constant-result-type-arg-p) - 'consed-sequence - (let ((result-type-arg-value - (continuation-value result-type-arg))) - (if (null result-type-arg-value) - 'null - result-type-arg-value))))) +(deftransform map ((result-type-arg fun seq &rest seqs) * * :node node) + (let* ((seq-names (make-gensym-list (1+ (length seqs)))) + (bare `(%map result-type-arg fun ,@seq-names)) + (constant-result-type-arg-p (constant-lvar-p result-type-arg)) + ;; what we know about the type of the result. (Note that the + ;; "result type" argument is not necessarily the type of the + ;; result, since NIL means the result has NULL type.) + (result-type (if (not constant-result-type-arg-p) + 'consed-sequence + (let ((result-type-arg-value + (lvar-value result-type-arg))) + (if (null result-type-arg-value) + 'null + result-type-arg-value))))) `(lambda (result-type-arg fun ,@seq-names) (truly-the ,result-type - ,(cond ((policy node (> speed safety)) - bare) - ((not constant-result-type-arg-p) - `(sequence-of-checked-length-given-type ,bare - result-type-arg)) - (t - (let ((result-ctype (specifier-type result-type))) - (if (array-type-p result-ctype) - (let* ((dims (array-type-dimensions result-ctype)) - (dim (first dims))) - (if (eq dim '*) - bare - `(vector-of-checked-length-given-length ,bare - ,dim))) - bare)))))))) + ,(cond ((policy node (< safety 3)) + ;; ANSI requires the length-related type check only + ;; when the SAFETY quality is 3... in other cases, we + ;; skip it, because it could be expensive. + bare) + ((not constant-result-type-arg-p) + `(sequence-of-checked-length-given-type ,bare + result-type-arg)) + (t + (let ((result-ctype (ir1-transform-specifier-type + result-type))) + (if (array-type-p result-ctype) + (let ((dims (array-type-dimensions result-ctype))) + (unless (and (listp dims) (= (length dims) 1)) + (give-up-ir1-transform "invalid sequence type")) + (let ((dim (first dims))) + (if (eq dim '*) + bare + `(vector-of-checked-length-given-length ,bare + ,dim)))) + ;; FIXME: this is wrong, as not all subtypes of + ;; VECTOR are ARRAY-TYPEs [consider, for + ;; example, (OR (VECTOR T 3) (VECTOR T + ;; 4))]. However, it's difficult to see what we + ;; should put here... maybe we should + ;; GIVE-UP-IR1-TRANSFORM if the type is a + ;; subtype of VECTOR but not an ARRAY-TYPE? + bare)))))))) + +;;; Return a DO loop, mapping a function FUN to elements of +;;; sequences. SEQS is a list of lvars, SEQ-NAMES - list of variables, +;;; bound to sequences, INTO - a variable, which is used in +;;; MAP-INTO. RESULT and BODY are forms, which can use variables +;;; FUNCALL-RESULT, containing the result of application of FUN, and +;;; INDEX, containing the current position in sequences. +(defun build-sequence-iterator (seqs seq-names &key result into body) + (declare (type list seqs seq-names) + (type symbol into)) + (collect ((bindings) + (declarations) + (vector-lengths) + (tests) + (places)) + (let ((found-vector-p nil)) + (flet ((process-vector (length) + (unless found-vector-p + (setq found-vector-p t) + (bindings `(index 0 (1+ index))) + (declarations `(type index index))) + (vector-lengths length))) + (loop for seq of-type lvar in seqs + for seq-name in seq-names + for type = (lvar-type seq) + do (cond ((csubtypep type (specifier-type 'list)) + (with-unique-names (index) + (bindings `(,index ,seq-name (cdr ,index))) + (declarations `(type list ,index)) + (places `(car ,index)) + (tests `(endp ,index)))) + ((csubtypep type (specifier-type 'vector)) + (process-vector `(length ,seq-name)) + (places `(locally (declare (optimize (insert-array-bounds-checks 0))) + (aref ,seq-name index)))) + (t + (give-up-ir1-transform + "can't determine sequence argument type")))) + (when into + (process-vector `(array-dimension ,into 0)))) + (when found-vector-p + (bindings `(length (min ,@(vector-lengths)))) + (tests `(>= index length))) + `(do (,@(bindings)) + ((or ,@(tests)) ,result) + (declare ,@(declarations)) + (let ((funcall-result (funcall fun ,@(places)))) + (declare (ignorable funcall-result)) + ,body))))) ;;; Try to compile %MAP efficiently when we can determine sequence ;;; argument types at compile time. @@ -115,124 +182,123 @@ ;;; handle that case more efficiently, but it's left as an exercise to ;;; the reader, because the code is complicated enough already and I ;;; don't happen to need that functionality right now. -- WHN 20000410 -(deftransform %map ((result-type fun &rest seqs) * * :policy (>= speed space)) +(deftransform %map ((result-type fun seq &rest seqs) * * + :policy (>= speed space)) "open code" - (unless seqs (abort-ir1-transform "no sequence args")) - (unless (constant-continuation-p result-type) + (unless (constant-lvar-p result-type) (give-up-ir1-transform "RESULT-TYPE argument not constant")) - (labels (;; 1-valued SUBTYPEP, fails unless second value of SUBTYPEP is true - (fn-1subtypep (fn x y) - (multiple-value-bind (subtype-p valid-p) (funcall fn x y) - (if valid-p - subtype-p - (give-up-ir1-transform - "can't analyze sequence type relationship")))) - (1subtypep (x y) (fn-1subtypep #'sb!xc:subtypep x y)) - (1csubtypep (x y) (fn-1subtypep #'csubtypep x y)) - (seq-supertype (seq) - (let ((ctype (continuation-type seq))) - (cond ((1csubtypep ctype (specifier-type 'vector)) 'vector) - ((1csubtypep ctype (specifier-type 'list)) 'list) - (t - (give-up-ir1-transform - "can't determine sequence argument type")))))) - (let* ((result-type-value (continuation-value result-type)) - (result-supertype (cond ((null result-type-value) 'null) - ((1subtypep result-type-value 'vector) - 'vector) - ((1subtypep result-type-value 'list) - 'list) - (t - (give-up-ir1-transform - "can't determine result type")))) - (seq-supertypes (mapcar #'seq-supertype seqs))) - (cond ((and result-type-value (= 1 (length seqs))) - ;; The consing arity-1 cases can be implemented - ;; reasonably efficiently as function calls, and the cost - ;; of consing should be significantly larger than - ;; function call overhead, so we always compile these - ;; cases as full calls regardless of speed-versus-space - ;; optimization policy. - (cond ((subtypep 'list result-type-value) - '(apply #'%map-to-list-arity-1 fun seqs)) - (;; (This one can be inefficient due to COERCE, but - ;; the current open-coded implementation has the - ;; same problem.) - (subtypep result-type-value 'vector) - `(coerce (apply #'%map-to-simple-vector-arity-1 fun seqs) - ',result-type-value)) - (t (bug "impossible (?) sequence type")))) - (t - (let* ((seq-args (make-gensym-list (length seqs))) - (index-bindingoids - (mapcar (lambda (seq-arg seq-supertype) - (let ((i (gensym "I"))) - (ecase seq-supertype - (vector `(,i 0 (1+ ,i))) - (list `(,i ,seq-arg (rest ,i)))))) - seq-args seq-supertypes)) - (indices (mapcar #'first index-bindingoids)) - (index-decls (mapcar (lambda (index seq-supertype) - `(type ,(ecase seq-supertype - (vector 'index) - (list 'list)) - ,index)) - indices seq-supertypes)) - (tests (mapcar (lambda (seq-arg seq-supertype index) - (ecase seq-supertype - (vector `(>= ,index (length ,seq-arg))) - (list `(endp ,index)))) - seq-args seq-supertypes indices)) - (values (mapcar (lambda (seq-arg seq-supertype index) - (ecase seq-supertype - (vector `(aref ,seq-arg ,index)) - (list `(first ,index)))) - seq-args seq-supertypes indices))) - (multiple-value-bind (push-dacc final-result) - (ecase result-supertype - (null (values nil nil)) - (list (values `(push dacc acc) `(nreverse acc))) - (vector (values `(push dacc acc) - `(coerce (nreverse acc) - ',result-type-value)))) - ;; (We use the same idiom, of returning a LAMBDA from - ;; DEFTRANSFORM, as is used in the DEFTRANSFORMs for - ;; FUNCALL and ALIEN-FUNCALL, and for the same - ;; reason: we need to get the runtime values of each - ;; of the &REST vars.) - `(lambda (result-type fun ,@seq-args) - (declare (ignore result-type)) - (do ((really-fun (%coerce-callable-to-fun fun)) - ,@index-bindingoids - (acc nil)) - ((or ,@tests) - ,final-result) - (declare ,@index-decls) - (declare (type list acc)) - (declare (ignorable acc)) - (let ((dacc (funcall really-fun ,@values))) - (declare (ignorable dacc)) - ,push-dacc)))))))))) + (labels ( ;; 1-valued SUBTYPEP, fails unless second value of SUBTYPEP is true + (fn-1subtypep (fn x y) + (multiple-value-bind (subtype-p valid-p) (funcall fn x y) + (if valid-p + subtype-p + (give-up-ir1-transform + "can't analyze sequence type relationship")))) + (1subtypep (x y) (fn-1subtypep #'sb!xc:subtypep x y))) + (let* ((result-type-value (lvar-value result-type)) + (result-supertype (cond ((null result-type-value) 'null) + ((1subtypep result-type-value 'vector) + 'vector) + ((1subtypep result-type-value 'list) + 'list) + (t + (give-up-ir1-transform + "result type unsuitable"))))) + (cond ((and result-type-value (null seqs)) + ;; The consing arity-1 cases can be implemented + ;; reasonably efficiently as function calls, and the cost + ;; of consing should be significantly larger than + ;; function call overhead, so we always compile these + ;; cases as full calls regardless of speed-versus-space + ;; optimization policy. + (cond ((subtypep result-type-value 'list) + '(%map-to-list-arity-1 fun seq)) + ( ;; (This one can be inefficient due to COERCE, but + ;; the current open-coded implementation has the + ;; same problem.) + (subtypep result-type-value 'vector) + `(coerce (%map-to-simple-vector-arity-1 fun seq) + ',result-type-value)) + (t (bug "impossible (?) sequence type")))) + (t + (let* ((seqs (cons seq seqs)) + (seq-args (make-gensym-list (length seqs)))) + (multiple-value-bind (push-dacc result) + (ecase result-supertype + (null (values nil nil)) + (list (values `(push funcall-result acc) + `(nreverse acc))) + (vector (values `(push funcall-result acc) + `(coerce (nreverse acc) + ',result-type-value)))) + ;; (We use the same idiom, of returning a LAMBDA from + ;; DEFTRANSFORM, as is used in the DEFTRANSFORMs for + ;; FUNCALL and ALIEN-FUNCALL, and for the same + ;; reason: we need to get the runtime values of each + ;; of the &REST vars.) + `(lambda (result-type fun ,@seq-args) + (declare (ignore result-type)) + (let ((fun (%coerce-callable-to-fun fun)) + (acc nil)) + (declare (type list acc)) + (declare (ignorable acc)) + ,(build-sequence-iterator + seqs seq-args + :result result + :body push-dacc)))))))))) + +;;; MAP-INTO +(deftransform map-into ((result fun &rest seqs) + (vector * &rest *) + *) + "open code" + (let ((seqs-names (mapcar (lambda (x) + (declare (ignore x)) + (gensym)) + seqs))) + `(lambda (result fun ,@seqs-names) + ,(build-sequence-iterator + seqs seqs-names + :result '(when (array-has-fill-pointer-p result) + (setf (fill-pointer result) index)) + :into 'result + :body '(locally (declare (optimize (insert-array-bounds-checks 0))) + (setf (aref result index) funcall-result))) + result))) + -(deftransform elt ((s i) ((simple-array * (*)) *) * :when :both) +;;; FIXME: once the confusion over doing transforms with known-complex +;;; arrays is over, we should also transform the calls to (AND (ARRAY +;;; * (*)) (NOT (SIMPLE-ARRAY * (*)))) objects. +(deftransform elt ((s i) ((simple-array * (*)) *) *) '(aref s i)) -(deftransform elt ((s i) (list *) * :when :both) +(deftransform elt ((s i) (list *) * :policy (< safety 3)) '(nth i s)) -(deftransform %setelt ((s i v) ((simple-array * (*)) * *) * :when :both) +(deftransform %setelt ((s i v) ((simple-array * (*)) * *) *) '(%aset s i v)) -(deftransform %setelt ((s i v) (list * *)) +(deftransform %setelt ((s i v) (list * *) * :policy (< safety 3)) '(setf (car (nthcdr i s)) v)) +(deftransform %check-vector-sequence-bounds ((vector start end) + (vector * *) * + :node node) + (if (policy node (< safety speed)) + '(or end (length vector)) + '(let ((length (length vector))) + (if (<= 0 start (or end length) length) + (or end length) + (sb!impl::signal-bounding-indices-bad-error vector start end))))) + (macrolet ((def (name) `(deftransform ,name ((e l &key (test #'eql)) * * - :node node :when :both) - (unless (constant-continuation-p l) + :node node) + (unless (constant-lvar-p l) (give-up-ir1-transform)) - (let ((val (continuation-value l))) + (let ((val (lvar-value l))) (unless (policy node (or (= speed 3) (and (>= speed space) @@ -266,9 +332,9 @@ ;; if ITEM is not a NUMBER or is a FIXNUM, apply ;; transform, else give up on transform. (cond (test - (unless (continuation-fun-is test '(eq)) + (unless (lvar-fun-is test '(eq)) (give-up-ir1-transform))) - ((types-equal-or-intersect (continuation-type item) + ((types-equal-or-intersect (lvar-type item) (specifier-type 'number)) (give-up-ir1-transform "Item might be a number."))) `(,',eq-fun item list)))) @@ -279,58 +345,63 @@ (deftransform delete-if ((pred list) (t list)) "open code" '(do ((x list (cdr x)) - (splice '())) + (splice '())) ((endp x) list) (cond ((funcall pred (car x)) - (if (null splice) - (setq list (cdr x)) - (rplacd splice (cdr x)))) - (T (setq splice x))))) + (if (null splice) + (setq list (cdr x)) + (rplacd splice (cdr x)))) + (t (setq splice x))))) (deftransform fill ((seq item &key (start 0) (end (length seq))) - (vector t &key (:start t) (:end index)) - * - :policy (> speed space)) + (vector t &key (:start t) (:end index)) + * + :policy (> speed space)) "open code" (let ((element-type (upgraded-element-type-specifier-or-give-up seq))) - `(with-array-data ((data seq) - (start start) - (end end)) + (values + `(with-array-data ((data seq) + (start start) + (end end)) (declare (type (simple-array ,element-type 1) data)) + (declare (type fixnum start end)) (do ((i start (1+ i))) - ((= i end) seq) - (declare (type index i)) - ;; WITH-ARRAY-DATA did our range checks once and for all, so - ;; it'd be wasteful to check again on every AREF. - (declare (optimize (safety 0))) - (setf (aref data i) item))))) + ((= i end) seq) + (declare (type index i)) + ;; WITH-ARRAY-DATA did our range checks once and for all, so + ;; it'd be wasteful to check again on every AREF... + (declare (optimize (safety 0))) + (setf (aref data i) item))) + ;; ... though we still need to check that the new element can fit + ;; into the vector in safe code. -- CSR, 2002-07-05 + `((declare (type ,element-type item)))))) ;;;; utilities -;;; Return true if CONT's only use is a non-NOTINLINE reference to a +;;; Return true if LVAR's only use is a non-NOTINLINE reference to a ;;; global function with one of the specified NAMES. -(defun continuation-fun-is (cont names) - (declare (type continuation cont) (list names)) - (let ((use (continuation-use cont))) +(defun lvar-fun-is (lvar names) + (declare (type lvar lvar) (list names)) + (let ((use (lvar-uses lvar))) (and (ref-p use) - (let ((leaf (ref-leaf use))) - (and (global-var-p leaf) - (eq (global-var-kind leaf) :global-function) - (not (null (member (leaf-source-name leaf) names - :test #'equal)))))))) - -;;; If CONT is a constant continuation, the return the constant value. -;;; If it is null, then return default, otherwise quietly give up the -;;; IR1 transform. + (let ((leaf (ref-leaf use))) + (and (global-var-p leaf) + (eq (global-var-kind leaf) :global-function) + (not (null (member (leaf-source-name leaf) names + :test #'equal)))))))) + +;;; If LVAR is a constant lvar, the return the constant value. If it +;;; is null, then return default, otherwise quietly give up the IR1 +;;; transform. ;;; ;;; ### Probably should take an ARG and flame using the NAME. -(defun constant-value-or-lose (cont &optional default) - (declare (type (or continuation null) cont)) - (cond ((not cont) default) - ((constant-continuation-p cont) - (continuation-value cont)) - (t - (give-up-ir1-transform)))) +(defun constant-value-or-lose (lvar &optional default) + (declare (type (or lvar null) lvar)) + (cond ((not lvar) default) + ((constant-lvar-p lvar) + (lvar-value lvar)) + (t + (give-up-ir1-transform)))) ;;; FIXME: Why is this code commented out? (Why *was* it commented ;;; out? We inherited this situation from cmucl-2.4.8, with no @@ -341,7 +412,7 @@ ;;; the argument (which should be referenced in any expansion), and ;;; the continuation for that argument (or NIL if unsupplied.) (defstruct (arg (:constructor %make-arg (name cont)) - (:copier nil)) + (:copier nil)) (name nil :type symbol) (cont nil :type (or continuation null))) (defmacro make-arg (name) @@ -362,10 +433,10 @@ (declare (type (or arg null) arg)) (if (and arg (arg-cont arg)) (let ((cont (arg-cont arg))) - (unless (constant-continuation-p cont) - (give-up-ir1-transform "Argument is not constant: ~S." - (arg-name arg))) - (continuation-value from-end)) + (unless (constant-continuation-p cont) + (give-up-ir1-transform "Argument is not constant: ~S." + (arg-name arg))) + (continuation-value from-end)) default)) ;;; If Arg is a constant and is EQL to X, then return T, otherwise NIL. If @@ -374,8 +445,8 @@ (declare (type (or arg null) x)) (if (and arg (arg-cont arg)) (let ((cont (arg-cont arg))) - (and (constant-continuation-p cont) - (eql (continuation-value cont) x))) + (and (constant-continuation-p cont) + (eql (continuation-value cont) x))) (eql default x))) (defstruct (iterator (:copier nil)) @@ -388,9 +459,9 @@ ;; A form that returns the current value. This may be set with SETF to set ;; the current value. (current (error "Must specify CURRENT.")) - ;; In a :Normal iterator, a form that tests whether there is a current value. + ;; In a :NORMAL iterator, a form that tests whether there is a current value. (done nil) - ;; In a :Result iterator, a form that truncates the result at the current + ;; In a :RESULT iterator, a form that truncates the result at the current ;; position and returns it. (result nil) ;; A form that returns the initial total number of values. The result is @@ -413,65 +484,65 @@ ;;; the iteration is forward or backward, then GIVE-UP. (defun make-sequence-iterator (sequence type &key start end from-end index) (declare (symbol sequence) (type ctype type) - (type (or arg null) start end from-end) - (type (or symbol null) index)) + (type (or arg null) start end from-end) + (type (or symbol null) index)) (let ((from-end (arg-constant-value from-end nil))) (cond ((csubtypep type (specifier-type 'vector)) - (let* ((n-stop (gensym)) - (n-idx (or index (gensym))) - (start (default-arg 0 start)) - (end (default-arg `(length ,sequence) end))) - (make-iterator - :kind :normal - :binds `((,n-idx ,(if from-end `(1- ,end) ,start)) - (,n-stop ,(if from-end `(1- ,start) ,end))) - :decls `((type neg-index ,n-idx ,n-stop)) - :current `(aref ,sequence ,n-idx) - :done `(,(if from-end '<= '>=) ,n-idx ,n-stop) - :next `(setq ,n-idx - ,(if from-end `(1- ,n-idx) `(1+ ,n-idx))) - :length (if from-end - `(- ,n-idx ,n-stop) - `(- ,n-stop ,n-idx))))) - ((csubtypep type (specifier-type 'list)) - (let* ((n-stop (if (and end (not from-end)) (gensym) nil)) - (n-current (gensym)) - (start-p (not (arg-eql start 0 0))) - (end-p (not (arg-eql end nil nil))) - (start (default-arg start 0)) - (end (default-arg end nil))) - (make-iterator - :binds `((,n-current - ,(if from-end - (if (or start-p end-p) - `(nreverse (subseq ,sequence ,start - ,@(when end `(,end)))) - `(reverse ,sequence)) - (if start-p - `(nthcdr ,start ,sequence) - sequence))) - ,@(when n-stop - `((,n-stop (nthcdr (the index - (- ,end ,start)) - ,n-current)))) - ,@(when index - `((,index ,(if from-end `(1- ,end) start))))) - :kind :normal - :decls `((list ,n-current ,n-end) - ,@(when index `((type neg-index ,index)))) - :current `(car ,n-current) - :done `(eq ,n-current ,n-stop) - :length `(- ,(or end `(length ,sequence)) ,start) - :next `(progn - (setq ,n-current (cdr ,n-current)) - ,@(when index - `((setq ,n-idx - ,(if from-end - `(1- ,index) - `(1+ ,index))))))))) - (t - (give-up-ir1-transform - "can't tell whether sequence is a list or a vector"))))) + (let* ((n-stop (gensym)) + (n-idx (or index (gensym))) + (start (default-arg 0 start)) + (end (default-arg `(length ,sequence) end))) + (make-iterator + :kind :normal + :binds `((,n-idx ,(if from-end `(1- ,end) ,start)) + (,n-stop ,(if from-end `(1- ,start) ,end))) + :decls `((type neg-index ,n-idx ,n-stop)) + :current `(aref ,sequence ,n-idx) + :done `(,(if from-end '<= '>=) ,n-idx ,n-stop) + :next `(setq ,n-idx + ,(if from-end `(1- ,n-idx) `(1+ ,n-idx))) + :length (if from-end + `(- ,n-idx ,n-stop) + `(- ,n-stop ,n-idx))))) + ((csubtypep type (specifier-type 'list)) + (let* ((n-stop (if (and end (not from-end)) (gensym) nil)) + (n-current (gensym)) + (start-p (not (arg-eql start 0 0))) + (end-p (not (arg-eql end nil nil))) + (start (default-arg start 0)) + (end (default-arg end nil))) + (make-iterator + :binds `((,n-current + ,(if from-end + (if (or start-p end-p) + `(nreverse (subseq ,sequence ,start + ,@(when end `(,end)))) + `(reverse ,sequence)) + (if start-p + `(nthcdr ,start ,sequence) + sequence))) + ,@(when n-stop + `((,n-stop (nthcdr (the index + (- ,end ,start)) + ,n-current)))) + ,@(when index + `((,index ,(if from-end `(1- ,end) start))))) + :kind :normal + :decls `((list ,n-current ,n-end) + ,@(when index `((type neg-index ,index)))) + :current `(car ,n-current) + :done `(eq ,n-current ,n-stop) + :length `(- ,(or end `(length ,sequence)) ,start) + :next `(progn + (setq ,n-current (cdr ,n-current)) + ,@(when index + `((setq ,n-idx + ,(if from-end + `(1- ,index) + `(1+ ,index))))))))) + (t + (give-up-ir1-transform + "can't tell whether sequence is a list or a vector"))))) ;;; Make an iterator used for constructing result sequences. Name is a ;;; variable to be bound to the result sequence. Type is the type of result @@ -487,32 +558,32 @@ #!+sb-doc "COERCE-FUNCTIONS ({(Name Fun-Arg Default)}*) Form*" (collect ((binds) - (defs)) + (defs)) (dolist (spec specs) `(let ((body (progn ,@body)) - (n-fun (arg-name ,(second spec))) - (fun-cont (arg-cont ,(second spec)))) - (cond ((not fun-cont) - `(macrolet ((,',(first spec) (&rest args) - `(,',',(third spec) ,@args))) - ,body)) - ((not (csubtypep (continuation-type fun-cont) - (specifier-type 'function))) - (when (policy *compiler-error-context* - (> speed inhibit-warnings)) - (compiler-note - "~S may not be a function, so must coerce at run-time." - n-fun)) - (once-only ((n-fun `(if (functionp ,n-fun) - ,n-fun - (symbol-function ,n-fun)))) - `(macrolet ((,',(first spec) (&rest args) - `(funcall ,',n-fun ,@args))) - ,body))) - (t - `(macrolet ((,',(first spec) (&rest args) - `(funcall ,',n-fun ,@args))) - ,body))))))) + (n-fun (arg-name ,(second spec))) + (fun-cont (arg-cont ,(second spec)))) + (cond ((not fun-cont) + `(macrolet ((,',(first spec) (&rest args) + `(,',',(third spec) ,@args))) + ,body)) + ((not (csubtypep (continuation-type fun-cont) + (specifier-type 'function))) + (when (policy *compiler-error-context* + (> speed inhibit-warnings)) + (compiler-notify + "~S may not be a function, so must coerce at run-time." + n-fun)) + (once-only ((n-fun `(if (functionp ,n-fun) + ,n-fun + (symbol-function ,n-fun)))) + `(macrolet ((,',(first spec) (&rest args) + `(funcall ,',n-fun ,@args))) + ,body))) + (t + `(macrolet ((,',(first spec) (&rest args) + `(funcall ,',n-fun ,@args))) + ,body))))))) ;;; Wrap code around the result of the body to define Name as a local macro ;;; that returns true when its arguments satisfy the test according to the Args @@ -522,8 +593,8 @@ `(let ((not-p (arg-cont ,test-not))) (when (and (arg-cont ,test) not-p) (abort-ir1-transform "Both ~S and ~S were supplied." - (arg-name ,test) - (arg-name ,test-not))) + (arg-name ,test) + (arg-name ,test-not))) (coerce-funs ((,name (if not-p ,test-not ,test) eql)) ,@body))) |# @@ -552,17 +623,19 @@ ;;; Return a form that tests the free variables STRING1 and STRING2 ;;; for the ordering relationship specified by LESSP and EQUALP. The ;;; start and end are also gotten from the environment. Both strings -;;; must be SIMPLE-STRINGs. +;;; must be SIMPLE-BASE-STRINGs. (macrolet ((def (name lessp equalp) `(deftransform ,name ((string1 string2 start1 end1 start2 end2) - (simple-string simple-string t t t t) *) + (simple-base-string simple-base-string t t t t) *) `(let* ((end1 (if (not end1) (length string1) end1)) (end2 (if (not end2) (length string2) end2)) (index (sb!impl::%sp-string-compare string1 start1 end1 string2 start2 end2))) (if index - (cond ((= index ,(if ',lessp 'end1 'end2)) index) - ((= index ,(if ',lessp 'end2 'end1)) nil) + (cond ((= index end1) + ,(if ',lessp 'index nil)) + ((= (+ index (- start2 start1)) end2) + ,(if ',lessp nil 'index)) ((,(if ',lessp 'char< 'char>) (schar string1 index) (schar string2 @@ -570,7 +643,7 @@ (+ index (truly-the fixnum (- start2 - start1)))))) + start1)))))) index) (t nil)) ,(if ',equalp 'end1 nil)))))) @@ -581,7 +654,7 @@ (macrolet ((def (name result-fun) `(deftransform ,name ((string1 string2 start1 end1 start2 end2) - (simple-string simple-string t t t t) *) + (simple-base-string simple-base-string t t t t) *) `(,',result-fun (sb!impl::%sp-string-compare string1 start1 (or end1 (length string1)) @@ -590,99 +663,391 @@ (def string/=* identity)) -;;;; string-only transforms for sequence functions -;;;; -;;;; Note: CMU CL had more of these, including transforms for -;;;; functions which cons. In SBCL, we've gotten rid of most of the -;;;; transforms for functions which cons, since our GC overhead is -;;;; sufficiently large that it doesn't seem worth it to try to -;;;; economize on function call overhead or on the overhead of runtime -;;;; type dispatch in AREF. The exception is CONCATENATE, since -;;;; a full call to CONCATENATE would have to look up the sequence -;;;; type, which can be really slow. -;;;; -;;;; FIXME: It would be nicer for these transforms to work for any -;;;; calls when all arguments are vectors with the same element type, -;;;; rather than restricting them to STRINGs only. - -;;; FIXME: Shouldn't we be testing for legality of -;;; * START1, START2, END1, and END2 indices? -;;; * size of copied string relative to destination string? -;;; (Either there should be tests conditional on SAFETY>=SPEED, or -;;; the transform should be conditional on SPEED>SAFETY.) +;;;; transforms for sequence functions + +;;; Moved here from generic/vm-tran.lisp to satisfy clisp. Only applies +;;; to vectors based on simple arrays. +(def!constant vector-data-bit-offset + (* sb!vm:vector-data-offset sb!vm:n-word-bits)) + +(eval-when (:compile-toplevel) +(defun valid-bit-bash-saetp-p (saetp) + ;; BIT-BASHing isn't allowed on simple vectors that contain pointers + (and (not (eq t (sb!vm:saetp-specifier saetp))) + ;; Disallowing (VECTOR NIL) also means that we won't transform + ;; sequence functions into bit-bashing code and we let the + ;; generic sequence functions signal errors if necessary. + (not (zerop (sb!vm:saetp-n-bits saetp))) + ;; Due to limitations with the current BIT-BASHing code, we can't + ;; BIT-BASH reliably on arrays whose element types are larger + ;; than the word size. + (<= (sb!vm:saetp-n-bits saetp) sb!vm:n-word-bits))) +) ; EVAL-WHEN + +;;; FIXME: In the copy loops below, we code the loops in a strange +;;; fashion: +;;; +;;; (do ((i (+ src-offset length) (1- i))) +;;; ((<= i 0) ...) +;;; (... (aref foo (1- i)) ...)) +;;; +;;; rather than the more natural (and seemingly more efficient): +;;; +;;; (do ((i (1- (+ src-offset length)) (1- i))) +;;; ((< i 0) ...) +;;; (... (aref foo i) ...)) +;;; +;;; (more efficient because we don't have to do the index adjusting on +;;; every iteration of the loop) +;;; +;;; We do this to avoid a suboptimality in SBCL's backend. In the +;;; latter case, the backend thinks I is a FIXNUM (which it is), but +;;; when used as an array index, the backend thinks I is a +;;; POSITIVE-FIXNUM (which it is). However, since the backend thinks of +;;; these as distinct storage classes, it cannot coerce a move from a +;;; FIXNUM TN to a POSITIVE-FIXNUM TN. The practical effect of this +;;; deficiency is that we have two extra moves and increased register +;;; pressure, which can lead to some spectacularly bad register +;;; allocation. (sub-FIXME: the register allocation even with the +;;; strangely written loops is not always excellent, either...). Doing +;;; it the first way, above, means that I is always thought of as a +;;; POSITIVE-FIXNUM and there are no issues. +;;; +;;; Besides, the *-WITH-OFFSET machinery will fold those index +;;; adjustments in the first version into the array addressing at no +;;; performance penalty! + +;;; This transform is critical to the performance of string streams. If +;;; you tweak it, make sure that you compare the disassembly, if not the +;;; performance of, the functions implementing string streams +;;; (e.g. SB!IMPL::STRING-OUCH). +(macrolet + ((define-replace-transforms () + (loop for saetp across sb!vm:*specialized-array-element-type-properties* + for sequence-type = `(simple-array ,(sb!vm:saetp-specifier saetp) (*)) + unless (= (sb!vm:saetp-typecode saetp) sb!vm::simple-array-nil-widetag) + collect + `(deftransform replace ((seq1 seq2 &key (start1 0) (start2 0) end1 end2) + (,sequence-type ,sequence-type &rest t) + ,sequence-type + :node node) + ,(cond + ((valid-bit-bash-saetp-p saetp) nil) + ;; If we're not bit-bashing, only allow cases where we + ;; can determine the order of copying up front. (There + ;; are actually more cases we can handle if we know the + ;; amount that we're copying, but this handles the + ;; common cases.) + (t '(unless (= (constant-value-or-lose start1 0) + (constant-value-or-lose start2 0)) + (give-up-ir1-transform)))) + `(let* ((len1 (length seq1)) + (len2 (length seq2)) + (end1 (or end1 len1)) + (end2 (or end2 len2)) + (replace-len1 (- end1 start1)) + (replace-len2 (- end2 start2))) + ,(unless (policy node (= safety 0)) + `(progn + (unless (<= 0 start1 end1 len1) + (sb!impl::signal-bounding-indices-bad-error seq1 start1 end1)) + (unless (<= 0 start2 end2 len2) + (sb!impl::signal-bounding-indices-bad-error seq2 start2 end2)))) + ,',(cond + ((valid-bit-bash-saetp-p saetp) + (let* ((n-element-bits (sb!vm:saetp-n-bits saetp)) + (bash-function (intern (format nil "UB~D-BASH-COPY" n-element-bits) + (find-package "SB!KERNEL")))) + `(funcall (function ,bash-function) seq2 start2 + seq1 start1 (min replace-len1 replace-len2)))) + (t + ;; We can expand the loop inline here because we + ;; would have given up the transform (see above) + ;; if we didn't have constant matching start + ;; indices. + '(do ((i start1 (1+ i)) + (end (+ start1 + (min replace-len1 replace-len2)))) + ((>= i end)) + (declare (optimize (insert-array-bounds-checks 0))) + (setf (aref seq1 i) (aref seq2 i))))) + seq1)) + into forms + finally (return `(progn ,@forms))))) + (define-replace-transforms)) + +;;; Expand simple cases of UB-BASH-COPY inline. "simple" is +;;; defined as those cases where we are doing word-aligned copies from +;;; both the source and the destination and we are copying from the same +;;; offset from both the source and the destination. (The last +;;; condition is there so we can determine the direction to copy at +;;; compile time rather than runtime. Remember that UB-BASH-COPY +;;; acts like memmove, not memcpy.) These conditions may seem rather +;;; restrictive, but they do catch common cases, like allocating a (* 2 +;;; N)-size buffer and blitting in the old N-size buffer in. + +(defun frob-bash-transform (src src-offset + dst dst-offset + length n-elems-per-word) + (declare (ignore src dst length)) + (let ((n-bits-per-elem (truncate sb!vm:n-word-bits n-elems-per-word))) + (multiple-value-bind (src-word src-elt) + (truncate (lvar-value src-offset) n-elems-per-word) + (multiple-value-bind (dst-word dst-elt) + (truncate (lvar-value dst-offset) n-elems-per-word) + ;; Avoid non-word aligned copies. + (unless (and (zerop src-elt) (zerop dst-elt)) + (give-up-ir1-transform)) + ;; Avoid copies where we would have to insert code for + ;; determining the direction of copying. + (unless (= src-word dst-word) + (give-up-ir1-transform)) + ;; FIXME: The cross-compiler doesn't optimize TRUNCATE properly, + ;; so we have to do its work here. + `(let ((end (+ ,src-word ,(if (= n-elems-per-word 1) + 'length + `(truncate (the index length) ,n-elems-per-word))))) + (declare (type index end)) + ;; Handle any bits at the end. + (when (logtest length (1- ,n-elems-per-word)) + (let* ((extra (mod length ,n-elems-per-word)) + ;; FIXME: The shift amount on this ASH is + ;; *always* negative, but the backend doesn't + ;; have a NEGATIVE-FIXNUM primitive type, so we + ;; wind up with a pile of code that tests the + ;; sign of the shift count prior to shifting when + ;; all we need is a simple negate and shift + ;; right. Yuck. + (mask (ash #.(1- (ash 1 sb!vm:n-word-bits)) + (* (- extra ,n-elems-per-word) + ,n-bits-per-elem)))) + (setf (sb!kernel:%vector-raw-bits dst end) + (logior + (logandc2 (sb!kernel:%vector-raw-bits dst end) + (ash mask + ,(ecase sb!c:*backend-byte-order* + (:little-endian 0) + (:big-endian `(* (- ,n-elems-per-word extra) + ,n-bits-per-elem))))) + (logand (sb!kernel:%vector-raw-bits src end) + (ash mask + ,(ecase sb!c:*backend-byte-order* + (:little-endian 0) + (:big-endian `(* (- ,n-elems-per-word extra) + ,n-bits-per-elem))))))))) + ;; Copy from the end to save a register. + (do ((i end (1- i))) + ((<= i ,src-word)) + (setf (sb!kernel:%vector-raw-bits dst (1- i)) + (sb!kernel:%vector-raw-bits src (1- i))))))))) + +#.(loop for i = 1 then (* i 2) + collect `(deftransform ,(intern (format nil "UB~D-BASH-COPY" i) + "SB!KERNEL") + ((src src-offset + dst dst-offset + length) + ((simple-unboxed-array (*)) + (constant-arg index) + (simple-unboxed-array (*)) + (constant-arg index) + index) + *) + (frob-bash-transform src src-offset + dst dst-offset length + ,(truncate sb!vm:n-word-bits i))) into forms + until (= i sb!vm:n-word-bits) + finally (return `(progn ,@forms))) + +;;; We expand copy loops inline in SUBSEQ and COPY-SEQ if we're copying +;;; arrays with elements of size >= the word size. We do this because +;;; we know the arrays cannot alias (one was just consed), therefore we +;;; can determine at compile time the direction to copy, and for +;;; word-sized elements, UB-BASH-COPY will do a bit of +;;; needless checking to figure out what's going on. The same +;;; considerations apply if we are copying elements larger than the word +;;; size, with the additional twist that doing it inline is likely to +;;; cons far less than calling REPLACE and letting generic code do the +;;; work. ;;; -;;; FIXME: Also, the transform should probably be dependent on -;;; SPEED>SPACE. -(deftransform replace ((string1 string2 &key (start1 0) (start2 0) - end1 end2) - (simple-string simple-string &rest t)) - `(locally - (declare (optimize (safety 0))) - (bit-bash-copy string2 - (the index - (+ (the index (* start2 sb!vm:n-byte-bits)) - ,vector-data-bit-offset)) - string1 - (the index - (+ (the index (* start1 sb!vm:n-byte-bits)) - ,vector-data-bit-offset)) - (the index - (* (min (the index (- (or end1 (length string1)) - start1)) - (the index (- (or end2 (length string2)) - start2))) - sb!vm:n-byte-bits))) - string1)) +;;; However, we do not do this for elements whose size is < than the +;;; word size because we don't want to deal with any alignment issues +;;; inline. The UB*-BASH-COPY transforms might fix things up later +;;; anyway. + +(defun maybe-expand-copy-loop-inline (src src-offset dst dst-offset length + element-type) + (let ((saetp (find-saetp element-type))) + (aver saetp) + (if (>= (sb!vm:saetp-n-bits saetp) sb!vm:n-word-bits) + (expand-aref-copy-loop src src-offset dst dst-offset length) + `(locally (declare (optimize (safety 0))) + (replace ,dst ,src :start1 ,dst-offset :start2 ,src-offset :end1 ,length))))) + +(defun expand-aref-copy-loop (src src-offset dst dst-offset length) + (if (eql src-offset dst-offset) + `(do ((i (+ ,src-offset ,length) (1- i))) + ((<= i ,src-offset)) + (declare (optimize (insert-array-bounds-checks 0))) + (setf (aref ,dst (1- i)) (aref ,src (1- i)))) + ;; KLUDGE: The compiler is not able to derive that (+ offset + ;; length) must be a fixnum, but arrives at (unsigned-byte 29). + ;; We, however, know it must be so, as by this point the bounds + ;; have already been checked. + `(do ((i (truly-the fixnum (+ ,src-offset ,length)) (1- i)) + (j (+ ,dst-offset ,length) (1- j))) + ((<= i ,src-offset)) + (declare (optimize (insert-array-bounds-checks 0)) + (type (integer 0 #.sb!xc:array-dimension-limit) j i)) + (setf (aref ,dst (1- j)) (aref ,src (1- i)))))) + +(deftransform subseq ((seq start &optional end) + ((or (simple-unboxed-array (*)) simple-vector) t &optional t) + * :node node) + (let ((array-type (lvar-type seq))) + (unless (array-type-p array-type) + (give-up-ir1-transform)) + (let ((element-type (type-specifier (array-type-specialized-element-type array-type)))) + `(let* ((length (length seq)) + (end (or end length))) + ,(unless (policy node (= safety 0)) + '(progn + (unless (<= 0 start end length) + (sb!impl::signal-bounding-indices-bad-error seq start end)))) + (let* ((size (- end start)) + (result (make-array size :element-type ',element-type))) + ,(maybe-expand-copy-loop-inline 'seq 'start 'result 0 'size element-type) + result))))) + +(deftransform copy-seq ((seq) ((or (simple-unboxed-array (*)) simple-vector)) *) + (let ((array-type (lvar-type seq))) + (unless (array-type-p array-type) + (give-up-ir1-transform)) + (let ((element-type (type-specifier (array-type-specialized-element-type array-type)))) + `(let* ((length (length seq)) + (result (make-array length :element-type ',element-type))) + ,(maybe-expand-copy-loop-inline 'seq 0 'result 0 'length element-type) + result)))) + +;;; FIXME: it really should be possible to take advantage of the +;;; macros used in code/seq.lisp here to avoid duplication of code, +;;; and enable even funkier transformations. +(deftransform search ((pattern text &key (start1 0) (start2 0) end1 end2 + (test #'eql) + (key #'identity) + from-end) + (vector vector &rest t) + * + :policy (> speed (max space safety))) + "open code" + (let ((from-end (when (lvar-p from-end) + (unless (constant-lvar-p from-end) + (give-up-ir1-transform ":FROM-END is not constant.")) + (lvar-value from-end))) + (keyp (lvar-p key)) + (testp (lvar-p test))) + `(block search + (let ((end1 (or end1 (length pattern))) + (end2 (or end2 (length text))) + ,@(when keyp + '((key (coerce key 'function)))) + ,@(when testp + '((test (coerce test 'function))))) + (declare (type index start1 start2 end1 end2)) + (do (,(if from-end + '(index2 (- end2 (- end1 start1)) (1- index2)) + '(index2 start2 (1+ index2)))) + (,(if from-end + '(< index2 start2) + '(>= index2 end2)) + nil) + ;; INDEX2 is FIXNUM, not an INDEX, as right before the loop + ;; terminates is hits -1 when :FROM-END is true and :START2 + ;; is 0. + (declare (type fixnum index2)) + (when (do ((index1 start1 (1+ index1)) + (index2 index2 (1+ index2))) + ((>= index1 end1) t) + (declare (type index index1 index2)) + ,@(unless from-end + '((when (= index2 end2) + (return-from search nil)))) + (unless (,@(if testp + '(funcall test) + '(eql)) + ,(if keyp + '(funcall key (aref pattern index1)) + '(aref pattern index1)) + ,(if keyp + '(funcall key (aref text index2)) + '(aref text index2))) + (return nil))) + (return index2))))))) ;;; FIXME: It seems as though it should be possible to make a DEFUN ;;; %CONCATENATE (with a DEFTRANSFORM to translate constant RTYPE to ;;; CTYPE before calling %CONCATENATE) which is comparably efficient, ;;; at least once DYNAMIC-EXTENT works. +;;; +;;; FIXME: currently KLUDGEed because of bug 188 +;;; +;;; FIXME: disabled for sb-unicode: probably want it back +#!-sb-unicode (deftransform concatenate ((rtype &rest sequences) - (t &rest simple-string) - simple-string) - (collect ((lets) - (forms) - (all-lengths) - (args)) - (dolist (seq sequences) - (declare (ignore seq)) - (let ((n-seq (gensym)) - (n-length (gensym))) - (args n-seq) - (lets `(,n-length (the index (* (length ,n-seq) sb!vm:n-byte-bits)))) - (all-lengths n-length) - (forms `(bit-bash-copy ,n-seq ,vector-data-bit-offset - res start - ,n-length)) - (forms `(setq start (+ start ,n-length))))) - `(lambda (rtype ,@(args)) - (declare (ignore rtype)) - (let* (,@(lets) - (res (make-string (truncate (the index (+ ,@(all-lengths))) - sb!vm:n-byte-bits))) - (start ,vector-data-bit-offset)) - (declare (type index start ,@(all-lengths))) - ,@(forms) - res)))) + (t &rest (or simple-base-string + (simple-array nil (*)))) + simple-base-string + :policy (< safety 3)) + (loop for rest-seqs on sequences + for n-seq = (gensym "N-SEQ") + for n-length = (gensym "N-LENGTH") + for start = 0 then next-start + for next-start = (gensym "NEXT-START") + collect n-seq into args + collect `(,n-length (length ,n-seq)) into lets + collect n-length into all-lengths + collect next-start into starts + collect `(if (and (typep ,n-seq '(simple-array nil (*))) + (> ,n-length 0)) + (error 'nil-array-accessed-error) + (#.(let* ((i (position 'character sb!kernel::*specialized-array-element-types*)) + (saetp (aref sb!vm:*specialized-array-element-type-properties* i)) + (n-bits (sb!vm:saetp-n-bits saetp))) + (intern (format nil "UB~D-BASH-COPY" n-bits) + "SB!KERNEL")) + ,n-seq 0 res ,start ,n-length)) + into forms + collect `(setq ,next-start (+ ,start ,n-length)) into forms + finally + (return + `(lambda (rtype ,@args) + (declare (ignore rtype)) + (let* (,@lets + (res (make-string (the index (+ ,@all-lengths)) + :element-type 'base-char))) + (declare (type index ,@all-lengths)) + (let (,@(mapcar (lambda (name) `(,name 0)) starts)) + (declare (type index ,@starts)) + ,@forms) + res))))) ;;;; CONS accessor DERIVE-TYPE optimizers (defoptimizer (car derive-type) ((cons)) - (let ((type (continuation-type cons)) - (null-type (specifier-type 'null))) + (let ((type (lvar-type cons)) + (null-type (specifier-type 'null))) (cond ((eq type null-type) - null-type) - ((cons-type-p type) - (cons-type-car-type type))))) + null-type) + ((cons-type-p type) + (cons-type-car-type type))))) (defoptimizer (cdr derive-type) ((cons)) - (let ((type (continuation-type cons)) - (null-type (specifier-type 'null))) + (let ((type (lvar-type cons)) + (null-type (specifier-type 'null))) (cond ((eq type null-type) - null-type) - ((cons-type-p type) - (cons-type-cdr-type type))))) + null-type) + ((cons-type-p type) + (cons-type-cdr-type type))))) ;;;; FIND, POSITION, and their -IF and -IF-NOT variants @@ -690,146 +1055,131 @@ ;;; %FIND-POSITION-IF only when %FIND-POSITION-IF has an inline ;;; expansion, so we factor out the condition into this function. (defun check-inlineability-of-find-position-if (sequence from-end) - (let ((ctype (continuation-type sequence))) + (let ((ctype (lvar-type sequence))) (cond ((csubtypep ctype (specifier-type 'vector)) - ;; It's not worth trying to inline vector code unless we - ;; know a fair amount about it at compile time. - (upgraded-element-type-specifier-or-give-up sequence) - (unless (constant-continuation-p from-end) - (give-up-ir1-transform - "FROM-END argument value not known at compile time"))) - ((csubtypep ctype (specifier-type 'list)) - ;; Inlining on lists is generally worthwhile. - ) - (t - (give-up-ir1-transform - "sequence type not known at compile time"))))) + ;; It's not worth trying to inline vector code unless we + ;; know a fair amount about it at compile time. + (upgraded-element-type-specifier-or-give-up sequence) + (unless (constant-lvar-p from-end) + (give-up-ir1-transform + "FROM-END argument value not known at compile time"))) + ((csubtypep ctype (specifier-type 'list)) + ;; Inlining on lists is generally worthwhile. + ) + (t + (give-up-ir1-transform + "sequence type not known at compile time"))))) ;;; %FIND-POSITION-IF and %FIND-POSITION-IF-NOT for LIST data (macrolet ((def (name condition) - `(deftransform ,name ((predicate sequence from-end start end key) - (function list t t t function) - * - :policy (> speed space) - :important t) - "expand inline" - `(let ((index 0) - (find nil) - (position nil)) - (declare (type index index)) - (dolist (i sequence (values find position)) - (let ((key-i (funcall key i))) - (when (and end (>= index end)) - (return (values find position))) - (when (>= index start) - (,',condition (funcall predicate key-i) - ;; This hack of dealing with non-NIL - ;; FROM-END for list data by iterating - ;; forward through the list and keeping - ;; track of the last time we found a match - ;; might be more screwy than what the user - ;; expects, but it seems to be allowed by - ;; the ANSI standard. (And if the user is - ;; screwy enough to ask for FROM-END - ;; behavior on list data, turnabout is - ;; fair play.) - ;; - ;; It's also not enormously efficient, - ;; calling PREDICATE and KEY more often - ;; than necessary; but all the - ;; alternatives seem to have their own - ;; efficiency problems. - (if from-end - (setf find i - position index) - (return (values i index)))))) - (incf index)))))) + `(deftransform ,name ((predicate sequence from-end start end key) + (function list t t t function) + * + :policy (> speed space)) + "expand inline" + `(let ((index 0) + (find nil) + (position nil)) + (declare (type index index)) + (dolist (i sequence + (if (and end (> end index)) + (sb!impl::signal-bounding-indices-bad-error + sequence start end) + (values find position))) + (let ((key-i (funcall key i))) + (when (and end (>= index end)) + (return (values find position))) + (when (>= index start) + (,',condition (funcall predicate key-i) + ;; This hack of dealing with non-NIL + ;; FROM-END for list data by iterating + ;; forward through the list and keeping + ;; track of the last time we found a match + ;; might be more screwy than what the user + ;; expects, but it seems to be allowed by + ;; the ANSI standard. (And if the user is + ;; screwy enough to ask for FROM-END + ;; behavior on list data, turnabout is + ;; fair play.) + ;; + ;; It's also not enormously efficient, + ;; calling PREDICATE and KEY more often + ;; than necessary; but all the + ;; alternatives seem to have their own + ;; efficiency problems. + (if from-end + (setf find i + position index) + (return (values i index)))))) + (incf index)))))) (def %find-position-if when) (def %find-position-if-not unless)) - + ;;; %FIND-POSITION for LIST data can be expanded into %FIND-POSITION-IF ;;; without loss of efficiency. (I.e., the optimizer should be able ;;; to straighten everything out.) (deftransform %find-position ((item sequence from-end start end key test) - (t list t t t t t) - * - :policy (> speed space) - :important t) + (t list t t t t t) + * + :policy (> speed space)) "expand inline" '(%find-position-if (let ((test-fun (%coerce-callable-to-fun test))) - ;; I'm having difficulty believing I'm - ;; reading it right, but as far as I can see, - ;; the only guidance that ANSI gives for the - ;; order of arguments to asymmetric tests is - ;; the character-set dependent example from - ;; the definition of FIND, - ;; (find #\d "here are some.." :test #'char>) - ;; => #\Space - ;; (In ASCII, we have (CHAR> #\d #\SPACE)=>T.) - ;; (Neither the POSITION definition page nor - ;; section 17.2 ("Rules about Test Functions") - ;; seem to consider the possibility of - ;; asymmetry.) - ;; - ;; So, judging from the example, we want to - ;; do (FUNCALL TEST-FUN ITEM I), because - ;; (FUNCALL #'CHAR> #\d #\SPACE)=>T. - ;; - ;; -- WHN (whose attention was drawn to it by - ;; Alexey Dejneka's bug report/fix) - (lambda (i) - (funcall test-fun item i))) - sequence - from-end - start - end - (%coerce-callable-to-fun key))) + ;; The order of arguments for asymmetric tests + ;; (e.g. #'<, as opposed to order-independent + ;; tests like #'=) is specified in the spec + ;; section 17.2.1 -- the O/Zi stuff there. + (lambda (i) + (funcall test-fun item i))) + sequence + from-end + start + end + (%coerce-callable-to-fun key))) ;;; The inline expansions for the VECTOR case are saved as macros so ;;; that we can share them between the DEFTRANSFORMs and the default ;;; cases in the DEFUNs. (This isn't needed for the LIST case, because ;;; the DEFTRANSFORMs for LIST are less choosy about when to expand.) (defun %find-position-or-find-position-if-vector-expansion (sequence-arg - from-end - start - end-arg - element - done-p-expr) - (let ((offset (gensym "OFFSET")) - (block (gensym "BLOCK")) - (index (gensym "INDEX")) - (n-sequence (gensym "N-SEQUENCE-")) - (sequence (gensym "SEQUENCE")) - (n-end (gensym "N-END-")) - (end (gensym "END-"))) + from-end + start + end-arg + element + done-p-expr) + (with-unique-names (offset block index n-sequence sequence n-end end) `(let ((,n-sequence ,sequence-arg) - (,n-end ,end-arg)) + (,n-end ,end-arg)) (with-array-data ((,sequence ,n-sequence :offset-var ,offset) - (,start ,start) - (,end (or ,n-end (length ,n-sequence)))) + (,start ,start) + (,end (%check-vector-sequence-bounds + ,n-sequence ,start ,n-end))) (block ,block - (macrolet ((maybe-return () - '(let ((,element (aref ,sequence ,index))) - (when ,done-p-expr - (return-from ,block - (values ,element - (- ,index ,offset))))))) - (if ,from-end - (loop for ,index - ;; (If we aren't fastidious about declaring that - ;; INDEX might be -1, then (FIND 1 #() :FROM-END T) - ;; can send us off into never-never land, since - ;; INDEX is initialized to -1.) - of-type index-or-minus-1 - from (1- ,end) downto ,start do - (maybe-return)) - (loop for ,index of-type index from ,start below ,end do - (maybe-return)))) - (values nil nil)))))) + (macrolet ((maybe-return () + ;; WITH-ARRAY-DATA has already performed bounds + ;; checking, so we can safely elide the checks + ;; in the inner loop. + '(let ((,element (locally (declare (optimize (insert-array-bounds-checks 0))) + (aref ,sequence ,index)))) + (when ,done-p-expr + (return-from ,block + (values ,element + (- ,index ,offset))))))) + (if ,from-end + (loop for ,index + ;; (If we aren't fastidious about declaring that + ;; INDEX might be -1, then (FIND 1 #() :FROM-END T) + ;; can send us off into never-never land, since + ;; INDEX is initialized to -1.) + of-type index-or-minus-1 + from (1- ,end) downto ,start do + (maybe-return)) + (loop for ,index of-type index from ,start below ,end do + (maybe-return)))) + (values nil nil)))))) (def!macro %find-position-vector-macro (item sequence - from-end start end key test) - (let ((element (gensym "ELEMENT"))) + from-end start end key test) + (with-unique-names (element) (%find-position-or-find-position-if-vector-expansion sequence from-end @@ -842,8 +1192,8 @@ `(funcall ,test ,item (funcall ,key ,element))))) (def!macro %find-position-if-vector-macro (predicate sequence - from-end start end key) - (let ((element (gensym "ELEMENT"))) + from-end start end key) + (with-unique-names (element) (%find-position-or-find-position-if-vector-expansion sequence from-end @@ -853,8 +1203,8 @@ `(funcall ,predicate (funcall ,key ,element))))) (def!macro %find-position-if-not-vector-macro (predicate sequence - from-end start end key) - (let ((element (gensym "ELEMENT"))) + from-end start end key) + (with-unique-names (element) (%find-position-or-find-position-if-vector-expansion sequence from-end @@ -866,31 +1216,112 @@ ;;; %FIND-POSITION, %FIND-POSITION-IF and %FIND-POSITION-IF-NOT for ;;; VECTOR data (deftransform %find-position-if ((predicate sequence from-end start end key) - (function vector t t t function) - * - :policy (> speed space) - :important t) + (function vector t t t function) + * + :policy (> speed space)) "expand inline" (check-inlineability-of-find-position-if sequence from-end) '(%find-position-if-vector-macro predicate sequence - from-end start end key)) + from-end start end key)) (deftransform %find-position-if-not ((predicate sequence from-end start end key) - (function vector t t t function) - * - :policy (> speed space) - :important t) + (function vector t t t function) + * + :policy (> speed space)) "expand inline" (check-inlineability-of-find-position-if sequence from-end) '(%find-position-if-not-vector-macro predicate sequence from-end start end key)) (deftransform %find-position ((item sequence from-end start end key test) - (t vector t t t function function) - * - :policy (> speed space) - :important t) + (t vector t t t function function) + * + :policy (> speed space)) "expand inline" (check-inlineability-of-find-position-if sequence from-end) '(%find-position-vector-macro item sequence - from-end start end key test)) + from-end start end key test)) + +;;; logic to unravel :TEST, :TEST-NOT, and :KEY options in FIND, +;;; POSITION-IF, etc. +(define-source-transform effective-find-position-test (test test-not) + (once-only ((test test) + (test-not test-not)) + `(cond + ((and ,test ,test-not) + (error "can't specify both :TEST and :TEST-NOT")) + (,test (%coerce-callable-to-fun ,test)) + (,test-not + ;; (Without DYNAMIC-EXTENT, this is potentially horribly + ;; inefficient, but since the TEST-NOT option is deprecated + ;; anyway, we don't care.) + (complement (%coerce-callable-to-fun ,test-not))) + (t #'eql)))) +(define-source-transform effective-find-position-key (key) + (once-only ((key key)) + `(if ,key + (%coerce-callable-to-fun ,key) + #'identity))) + +(macrolet ((define-find-position (fun-name values-index) + `(deftransform ,fun-name ((item sequence &key + from-end (start 0) end + key test test-not) + (t (or list vector) &rest t)) + '(nth-value ,values-index + (%find-position item sequence + from-end start + end + (effective-find-position-key key) + (effective-find-position-test + test test-not)))))) + (define-find-position find 0) + (define-find-position position 1)) + +(macrolet ((define-find-position-if (fun-name values-index) + `(deftransform ,fun-name ((predicate sequence &key + from-end (start 0) + end key) + (t (or list vector) &rest t)) + '(nth-value + ,values-index + (%find-position-if (%coerce-callable-to-fun predicate) + sequence from-end + start end + (effective-find-position-key key)))))) + (define-find-position-if find-if 0) + (define-find-position-if position-if 1)) + +;;; the deprecated functions FIND-IF-NOT and POSITION-IF-NOT. We +;;; didn't bother to worry about optimizing them, except note that on +;;; Sat, Oct 06, 2001 at 04:22:38PM +0100, Christophe Rhodes wrote on +;;; sbcl-devel +;;; +;;; My understanding is that while the :test-not argument is +;;; deprecated in favour of :test (complement #'foo) because of +;;; semantic difficulties (what happens if both :test and :test-not +;;; are supplied, etc) the -if-not variants, while officially +;;; deprecated, would be undeprecated were X3J13 actually to produce +;;; a revised standard, as there are perfectly legitimate idiomatic +;;; reasons for allowing the -if-not versions equal status, +;;; particularly remove-if-not (== filter). +;;; +;;; This is only an informal understanding, I grant you, but +;;; perhaps it's worth optimizing the -if-not versions in the same +;;; way as the others? +;;; +;;; FIXME: Maybe remove uses of these deprecated functions within the +;;; implementation of SBCL. +(macrolet ((define-find-position-if-not (fun-name values-index) + `(deftransform ,fun-name ((predicate sequence &key + from-end (start 0) + end key) + (t (or list vector) &rest t)) + '(nth-value + ,values-index + (%find-position-if-not (%coerce-callable-to-fun predicate) + sequence from-end + start end + (effective-find-position-key key)))))) + (define-find-position-if-not find-if-not 0) + (define-find-position-if-not position-if-not 1))